Liquid AI launches LFM3: Its 1.7B small model outperforms 9B models in running web agents.

Beating AI Express News: Liquid AI has launched its third-generation base model LFM3, rolling out six small models each with around 2 billion parameters, covering text, vision, speech and decision-making tasks. The 1.7B text model supports a 256K context window, with 4-bit quantized weights weighing only about 1GB. All six models support single-GPU fine-tuning, allowing enterprises to train them on their own data before deploying to mobile devices, personal computers or private servers. Liquid AI redesigned the model architecture, combining short convolutions, local attention and global attention, and specialized training for agent tasks such as tool calls and web operations. Compared to the previous LFM2.5-2.6B, the new models have around 35% fewer parameters but deliver significant performance improvements on some tasks. In official tests, LFM3-1.7B scored 38.42 on the WebArena-Verified browser operation task, versus just 3.02 for the prior generation, and outperformed Qwen3.5-9B (28.94), which has roughly 5 times more parameters. In the AutomationBench enterprise workflow test, its task success rate reached 41.29%, while Qwen3.5-9B scored 15.10%. However, results reversed for complex programming and long text tasks: in the SWE-bench Verified programming test, LFM3-1.7B scored 43.4, lower than Qwen3.5-9B's 51.2. In the 128K long text test, the two models notched scores of 65.6 and 87.1 respectively. Liquid AI also explicitly noted in its model card that this small model is not recommended for complex agent programming tasks.